An AI SOP template turns a recurring task into a fill-in prompt anyone on the team can run: a role and task, context, inputs as {{variables}}, steps, 3 to 5 real examples, an output format, quality checks, and an owner with a review date. Store one master copy where people already work, and review it monthly.
Key takeaways
- A prompt becomes a team SOP when it has a fixed structure, real examples and a named owner.
- Start with two or three tasks that happen weekly, involve more than one person and produce a similar output every time.
- Use 3 to 5 varied examples from real approved work, including one edge case, wrapped in example tags.
- Keep one master copy in a shared library and run it from a Project or custom GPT, never from an old chat.
- Review each SOP monthly against real inputs, change one thing at a time, and record the date and version.
Why team prompts stop getting reused
Most small marketing teams already have good prompts. They live in someone's chat history, a Slack thread from three months ago, or a note on one person's laptop. The person who wrote it gets great results. Everyone else gets something close, rewrites half of it, and quietly stops using it.
An AI SOP fixes that. It is a standard operating procedure for one recurring task, written as a prompt anyone on the team can fill in and run. This post gives you an AI SOP template you can copy today, shows how to write the examples that make it work, and covers where to store SOPs and how to keep them current.
The gap between a prompt and an SOP comes down to three things: a fixed structure, real examples, and an owner. Without those, a prompt is a personal habit. With them, it is team process.
Pick the work that deserves an SOP
Not every task needs one. Writing and maintaining an SOP takes time, so spend it where the work repeats. Use this table to sort your list.
| Signal | Write an SOP | Skip it for now |
|---|---|---|
| How often | Weekly, or once per client per month | Once a quarter or less |
| Who does it | Two or more people | One specialist |
| Output shape | Similar every time | Different every time |
| Cost of a miss | A redo or an awkward client email | Low |
| Inputs | Clear: notes, data, a brief | Mostly in someone's head |
Good first candidates in a marketing team: client report summaries, meeting notes turned into action items, blog briefs from a keyword, ad copy variations from a brief, and social posts from a published article. Start with two or three. If you want a structured way to find them, the steps in how to run an AI workflow audit work well for this.
The AI SOP template (copy and paste)
This is the skeleton I use. The top block is for humans. Everything from the first tag down gets pasted into Claude or ChatGPT. Anything in double curly braces is a blank to fill in each time.
AI SOP: {{Name of the task}}
Owner: {{Name}} | Last reviewed: {{Date}} | Version: {{1.0}}
Use when: {{The trigger, e.g. monthly report data is in}}
Do not use when: {{Cases this SOP does not cover}}
Known issues: {{Anything the output still gets wrong}}
<role_and_task>
You are helping {{team or agency name}} {{do the task}} for {{audience}}.
The goal is {{the outcome, in one sentence}}.
</role_and_task>
<context>
{{Background the AI needs: who the client is, brand rules,
what good looks like, words or claims to avoid.}}
Tone: {{e.g. friendly, professional, plain}}
</context>
<inputs>
Client: {{CLIENT_NAME}}
Source material: {{PASTE_NOTES_DATA_OR_BRIEF}}
Audience: {{WHO_WILL_READ_THIS}}
Length: {{TARGET_LENGTH}}
</inputs>
<instructions>
1. {{First step}}
2. {{Second step}}
3. {{Third step}}
4. If an input is missing or unclear, ask before writing.
</instructions>
<examples>
<example>
Input: {{A real past input}}
Output: {{The approved final version}}
</example>
<example>
Input: {{A different real input}}
Output: {{Approved output}}
</example>
<example>
Input: {{An edge case, e.g. bad month or missing data}}
Output: {{How we handled it}}
</example>
</examples>
<output_format>
{{Headings, bullets, word count, file format, what goes first}}
</output_format>
<quality_checks>
Before you answer, confirm:
- Every number matches the source material.
- Nothing is claimed that the inputs do not support.
- The tone matches the context section.
- {{One check specific to this task}}
</quality_checks>How each part earns its place
- Role and task. One or two sentences. OpenAI's guidance says prompts should be clear, specific and include enough context, and avoid ambiguity. Most failed prompts skip the goal.
- Context. Name the tone in plain words. OpenAI recommends describing tone with words like formal, friendly or professional.
- Inputs as blanks. In my experience, the fill-in blanks are what make people reuse an SOP. Nobody has to decide where their notes go.
- Examples in tags. Anthropic recommends putting each example in example tags, and the set in examples tags, so the model can tell examples apart from instructions.
- Quality checks. These catch the misses that cost you a redo: wrong numbers and claims nobody approved.
- Owner and date. Without them, nobody knows if the SOP is still right.
A filled-in snippet (illustrative)
Here is what the inputs block might look like for a monthly SEO summary. The client and figures are made up.
<inputs>
Client: Harbor Dental Group
Source material: Organic sessions 4,210 (last month 3,880). Top page: /invisalign. Two new referring domains.
Audience: Practice owner, not technical
Length: Under 200 words
</inputs>Write examples that teach the pattern
Examples do more work than any other part of the SOP. Anthropic's docs recommend 3 to 5 diverse, relevant examples and say they reduce misreadings of instructions, keep structure and style consistent, and help most on complex tasks.
The same company's best-practices page adds that examples should match your real use case, cover edge cases and vary enough that the model does not copy patterns you did not intend. If all three examples open with the same sentence, every output will too.
How to build a good set:
- Pull from approved work. Use outputs a client or lead actually signed off on, not drafts.
- Vary them. Different clients, different lengths, a good month and a flat month.
- Include one edge case. Missing data, a bad result, or an odd request, plus how you handled it.
- Strip confidential details. Swap names and figures if the SOP will be shared widely.
- Have the model review them. Anthropic suggests asking Claude to check your examples for relevance, variety and clarity, or to write more from your first few. I make this part of every monthly review.
Where to store prompt SOPs so people use them
Storage is where most prompt libraries die. The fix is to separate where the SOP lives from where it runs.
| Layer | What goes there | Who edits |
|---|---|---|
| Master library | One doc or folder per SOP, with version, owner and change log | The owner only |
| Where it runs | A shared Project or custom GPT in your workspace, with the SOP in its instructions or files | One or two people |
| Automation | A Make scenario, once the SOP is stable and the inputs arrive the same way every time | Whoever maintains your automations |
A few rules that hold up in practice:
- One master copy. If the SOP exists in five places, four of them are out of date. Everything else links back to the master.
- Never store an SOP in an old chat. Nobody can find it, and a new chat does not start from it. Put the SOP in the Project's instructions or files, where your plan supports that.
- Keep edit rights narrow. In my experience, one or two editors per SOP is the right number. Everyone else uses it and sends suggestions.
- Name things the same way. Something like "SOP | Monthly report summary | v1.3" makes the library searchable.
- Put it where work starts. Link the SOP from the project management task or client folder where the work begins, not from a wiki page nobody opens.
If you are setting up shared Projects in Claude or ChatGPT Business for the first time, the walkthrough in ChatGPT brand voice for teams covers the setup. Brand rules belong at the Project level; task SOPs sit on top.
Keep your AI SOP template current with a monthly review
An SOP that worked in January can drift by June. Clients change, your offer changes, and the models change. OpenAI describes prompting as an iterative process: start with a prompt, review the response, then adjust the wording, add context or simplify.
The good news for a small team: Anthropic notes that prompt changes show results almost immediately, while fine-tuning can take hours or days. A stale SOP is an afternoon fix, not a project.
Copy this monthly checklist for each SOP owner:
- Collect flags. Anyone who had to rewrite more than a few lines of output drops a note in one shared channel or doc.
- Run the SOP on last month's real inputs.
- Compare the output to what was approved and sent.
- Change one thing at a time, then rerun. When a prompt is not working, Anthropic suggests trying techniques in order from most to least broadly effective, starting with being clear and direct, then examples.
- Refresh at least one example with recent approved work.
- Update the version, the last-reviewed date and the change log.
- Retire any SOP nobody used last month, or find out why.
Roll it out so the team uses it
A library nobody opens is a document, not a process. Show each new SOP in a team meeting, run it live on a real task, and ask one person to use it for the next week and report back. Track use loosely: if the SOP is not being opened, the problem is usually storage or the inputs block, not the team.
For a fuller rollout plan, see how to train your team on AI in 30 days. Once an SOP runs the same way every month, it is a candidate for automation. The client reporting setup with Make and Claude is a working example of that next step.
Do it yourself or get help
You can do all of this in-house. Pick two tasks, copy the template above, gather three approved examples for each, and put them in a shared Project with one owner. Most teams can have a first SOP running in a week.
Where teams tend to want help is the first step and the upkeep: deciding which workflows are worth an SOP, and keeping a growing library and its automations working. That is what Keeplight does. The AI Workflow Audit looks at every recurring workflow and tells you what to build first and what to leave alone; you can see what the report looks like in the sample audit. Brand Brain Setup puts your voice, examples and rules into your team's AI so every SOP starts from the same base. If you want to talk it through, a free 20-minute intro call is the easiest place to start.
Frequently asked questions
What is an AI SOP?
An AI SOP is a standard operating procedure for one recurring task, written as a reusable prompt. It includes a fixed structure, fill-in blanks for inputs, real examples, an output format and an owner. The point is that anyone on the team gets the same quality, not only the person who wrote it.
How many examples should I put in a prompt SOP?
Anthropic recommends 3 to 5 diverse, relevant examples. Pull them from approved past work, vary them across clients and situations, and include at least one edge case such as missing data or a bad month.
Where should my team store shared prompts?
Keep one master copy in a shared library doc or folder with an owner and version number. Run it from a shared Project or custom GPT in your workspace, and link to it from where the task starts. Avoid storing SOPs in old chats, where nobody can find them.
How often should we update our AI SOPs?
Monthly works for most small teams. Run each SOP on last month's real inputs, compare against what was approved, change one thing at a time, and refresh at least one example. Update the version and review date every time.
Does the same AI SOP template work in Claude and ChatGPT?
Yes, the structure carries over. Clear tasks, context, examples and an output format help in both tools. The tags in the template come from Anthropic's guidance for Claude, and they do no harm in ChatGPT as plain labels.
Should we automate a prompt SOP?
Only once it is stable and the inputs arrive the same way every time. Run it manually for a few weeks first. When the output rarely needs edits, a tool like Make can feed it inputs and deliver results on a schedule.
Bryce Walls, Founder, Keeplight
Bryce builds the AI systems he runs his own marketing on. He runs campaigns, measures them, and automates the checks himself, and Keeplight sets up the same kind of AI and automation for small marketing teams.